End-to-end multimodal attention fusion for drug-target interaction prediction with cold-scaffold validation on the...

Predicting drug-target interactions computationally is a practical strategy for prioritizing candidate compounds in early drug discovery, but the reliability of published models is often limited by small resampled datasets, warm-only evaluation protocols, and architectures that prevent joint optimization of all model components. This study presents the Multimodal Attention Fusion Network (MAFN), an end-to-end PyTorch model that integrates 2,058-dimensional drug features, comprising ECFP4 Morgan fingerprints and ten physicochemical descriptors, with 567-dimensional protein sequence features, comprising amino acid composition, dipeptide composition, and CTD descriptors, through a jointly optimized attention-based fusion layer. The model is trained and evaluated on the Davis kinase-inhibitor benchmark: 25,772 unique drug-target pairs across 68 inhibitors and 379 kinase targets with zero pair duplication. Three evaluation protocols were applied: a stratified random 80/20 hold-out, Cold-Drug Leave-One-Out validation across 88 folds covering three random seeds, and Cold-Target Leave-One-Out across 86 folds. Two feature-matched Random Forest baselines, RF-ECFP4 and RF-CTD + FP, were evaluated on identical splits. Performance was additionally assessed at pKd thresholds of 6.5, 7.0, and 7.5 to confirm threshold independence. On the warm hold-out, MAFN achieved AUC-ROC 0.920, AUPR 0.595, F1-score 0.569, and MCC 0.535. Cold-Drug LOO yielded a median AUC of 0.762 (IQR 0.674 to 0.922, Cohen's d = 1.31) and Cold-Target LOO yielded 0.937 (IQR 0.861 to 0.985, d = 4.56). Notably, RF-ECFP4 outperformed MAFN under Cold-Drug LOO (median 0.924 vs 0.762), a finding reported and interpreted transparently. Permutation importance at the fused 512-dimensional embedding layer gave Spearman rho = -0.126 (p = 0.385). Source code and model checkpoints are provided as supplementary materials.